...

What we need is a generic range adaptor that can make a container of containers look like a "flat" sequence.

I suggest the syntax

rngOfContainers | boost::adaptors::flattened

and the underlying iterator should be up to bidirectional (if the underlying two container types support it).

Anyone interested in implementing this?

I've been interested in implementing something similar for a while. I was thinking slightly more generally that any n-dimensional structure can often be linearised or projected into a linear sequence. Hence this would enable not just the above example to work, but many standard algorithms could work with tree structures, hyper-cube trees, directed acyclic graphs etc. I've implemented a number of these already, but wanted to gain experience with them in my own projects before committing them to the trunk.

-Thorsten

Regards, Neil Groves

On 06/04/2011 23:14, Thorsten Ottosen wrote:

Den 06-04-2011 22:55, Neil Groves skrev:

...

I've been interested in implementing something similar for a while. I was thinking slightly more generally that any n-dimensional structure can often be linearised or projected into a linear sequence. Hence this would enable not just the above example to work, but many standard algorithms could work with tree structures, hyper-cube trees, directed acyclic graphs etc. I've implemented a number of these already, but wanted to gain experience with them in my own projects before committing them to the trunk.

Cool. I have a little problem seeing how it works with acyclic graphs. For example, there are many ways to do this. Which one do we pick?

True, there are many ways to linearize a n-dimensional structure. Consider a matrix, do you start with rows or with columns?

I think the container of containers case meets 90% or more of all use cases. A container of containers of containers can be done manually by applying it twice.

For ranges of ranges (of ranges...), unless they're random access, I don't think anything but linearizing in the "natural" order of iteration makes sense.

Den 07-04-2011 11:37, Mathias Gaunard skrev:

On 06/04/2011 19:55, Thorsten Ottosen wrote:

...

I suggest the syntax

rngOfContainers | boost::adaptors::flattened

and the underlying iterator should be up to bidirectional (if the underlying two container types support it).

Anyone interested in implementing this?

I have an implementation of this somewhere.

One annoying thing is that you cannot detect that a type is a range reliably.

Well, that is only a problem if you want to avoid manually applying the adaptor twice (or more), right? If so, I think that is a limitation we can live with. A simple solution covers most cases IMO. -Thorsten